| name | paude-spec |
| description | Write a task spec file for a paude container agent. Use when delegating work to paude, preparing a paude prompt-file, writing fire-and-forget agent tasks, or scoping submodule work for harvest. |
| argument-hint | [--project <name>] [--slug <slug>] [--output <absolute-path>] |
| allowed-tools | Read Write Shell Glob Grep |
Paude Spec
Produce a task spec that a container agent can execute autonomously and hand off for harvest without ambiguous commits or wrong submodule discipline.
<essential_principles>
- Spec states outcomes, not step-by-step implementation (unless safety-critical).
- Every spec names
session-name, harvest/<slug>, and repos likely touched.
- Spec ends with harvest-prep instructions — uncommitted work is lost.
- Use absolute output paths when the spec will be passed to
--prompt-file.
</essential_principles>
- Template: `templates/task-spec.md`
- Harvest prep (for container): `templates/harvest-prep-prompt.md`
- Also copied to: `devops/paude/harvest-prep-prompt.md` for human visibility
- Launch: `.agents/skills/paude-launch/SKILL.md`
- Planning specs: `.planning/paude-integration/task-specs/`
1. Gather intent
From conversation and $ARGUMENTS:
- Task goal and boundaries
--project <name> — planning directory (default: infer or paude-integration)
--slug <slug> — kebab-case task id
--output <path> — override default path
If scope is unclear, ask one sharp question before writing.
2. Choose output path
Default:
.planning/<project>/task-specs/<slug>.md
3. Write spec
Copy structure from templates/task-spec.md. Fill all sections.
Required fields:
- Session name (match planned
paude create name)
- Harvest branch
harvest/<slug>
- Repos table
- Commit discipline block
- "When finished" → harvest prep template reference
4. Copy harvest-prep for visibility
Ensure devops/paude/harvest-prep-prompt.md exists and matches templates/harvest-prep-prompt.md (update if drift).
5. Present to user
Show path and suggest:
paude-launch <session-name> --slug <slug> --spec <absolute-path>
<success_criteria>
- Spec file exists at agreed path
- Session name and harvest branch are explicit
- Submodule defaults and no-push rule included
- User has launch command with absolute spec path
</success_criteria>